An efficient detection model based on improved YOLOv5s for abnormal surface features of fish

Area of Science:

• Aquaculture
• Computer Vision
• Machine Learning

Background:

• Accurate detection of abnormal fish surface features is crucial for identifying aquatic health issues.
• Current methods suffer from subjectivity, low accuracy, and poor real-time performance.

Purpose of the Study:

• To develop a real-time, accurate fish surface abnormality detection model.
• To address limitations of existing fish health monitoring techniques.

Main Methods:

• An improved YOLOv5s model incorporating optimized Complete Intersection over Union (CIoU) and Non-Maximum Suppression (NMS) using normalized Gaussian Wasserstein distance for tiny target detection.
• Integration of the DenseOne module for feature reusability and MobileViTv2 for enhanced detection speed within the feature extraction network.
• Fusion of omni-dimensional dynamic convolution and convolutional block attention module based on ACmix principles for deep feature extraction in complex backgrounds.

Main Results:

• The model achieved 99.5% precision, 99.1% recall, 99.1% mAP50, 73.9% mAP50:95, and 88 FPS on 160 validation sets.
• Performance improvements over the baseline included +1.4% precision, +1.2% recall, +3.2% mAP50, +8.2% mAP50:95, and +1 FPS.
• The enhanced model demonstrated superior performance compared to other state-of-the-art models.

Conclusions:

• The proposed improved YOLOv5s model offers a significant advancement in real-time and accurate detection of abnormal fish surface features.
• The model's enhancements effectively address challenges related to subjectivity, accuracy, and speed in aquatic health assessment.
• This approach provides a robust tool for improving fish health monitoring and aquaculture management.